Providing adequate instruction and data bandwidth is a key problem in modern computer systems. In a conventional scalar architecture, each arithmetic operation, e.g., an addition or multiplication, requires one word of instruction bandwidth to control the operation and three words of data bandwidth to provide the input data and to consume the result (two words for the operands and one word for the result). Thus the raw bandwidth demand is four words per operation. Conventional architectures use a storage hierarchy consisting of register files and cache memories to provide much of this bandwidth; however, since arithmetic bandwidth scales with advances in technology, providing this instruction and data bandwidth at each level of the memory hierarchy, particularly the bottom, is a challenging problem.
Vector architectures have emerged as one approach to reducing the instruction bandwidth required for a computation. With conventional vector architectures, e.g., the Cray-1, a single instruction word specifies a sequence of arithmetic operations, one on each element of a vector of inputs. For example, a vector addition instruction VADD VA, VB, VC causes each element of an, e.g., sixty-four element vector VA to be added to the corresponding element of a vector VB with the result being placed in the corresponding element of vector VC. Thus, to the extent that the computation being performed can be expressed in terms of vector operations, a vector architecture reduces the required instruction bandwidth by a factor of the vector length (sixty-four in the case of the Cray-1).
While vector architectures may alleviate some of the instruction bandwidth requirements, data bandwidth demands remain undiminished. Each arithmetic operation still requires three words of data bandwidth from a global storage source shared by all arithmetic units. In most vector architectures, this global storage resource is the vector register file. As the number of arithmetic units is increased, this register file becomes a bottleneck that limits further improvements in machine performance.
To reduce the latency of arithmetic operations, some vector architectures perform “chaining” of arithmetic operations. For example, consider performing the above vector addition operation and then performing the vector multiplication operation VMUL VC VD VE using the result. With chaining, the vector multiply instruction consumes the elements computed by the vector add instruction in VC as they are produced and without waiting for the entire vector add instruction to complete. Chaining, however, also does not diminish the demand for data bandwidth—each arithmetic operation still requires three words of bandwidth from the vector register file.
Another latency problem arises in connection with conditional operations, i.e., operations in which the result is dependent on the result of a Boolean or multi-valued test on input data. For example, when sorting several values, two values are compared and, depending on whether the first is greater than, less than or equal to the second value, different actions may be taken.
As another example, consider chroma-keying a video signal. Chroma-keying is used to, e.g., superimpose one video stream representing a foreground object such as a television weather person on another video stream representing a background object such as a map. The foreground object is typically photographed against a blue or other fixed color background to facilitate separation of the object from its background based on color or chrominance. Using a C-like pseudocode, this process can be described by
for each pixel p[i] {read foreground pixel pf[i] from foreground stream;read background pixel pb[i] from background stream;if (pf[i] is blue) {p[i] = pb[i];do background processing; }else {p[i] = pf[i];do foreground processing; }output p[i] to output stream; }Since subsequent program execution may involve completely different data or completely different operations depending on the outcome of the comparison, execution generally halts until the result of the conditional operation is known, thereby serializing the program flow and lowering the performance of parallel processing systems.
In the above example, processing will proceed (using parallel operations if supported by the processor) until it encounters the conditional portion of the if-else statement, at which time it stops and waits for the conditional expression to be evaluated. The time, e.g., in clock cycles, from the time the condition is tested until the first instruction at the chosen branch destination is executed is called the branch latency of the instruction. Contemporary pipelined processors typically have a branch latency of about four cycles.
As noted above, during the branch latency period all functional units of the processor are idle. Since modern processors often have multiple functional units, the number of wasted processor cycles can be multiplied several times over, and this problem can be compounded by pipelining, another feature common to most modern microprocessors. In a pipelined processor having five functional units, for example, twenty instruction issue opportunities are lost to a conditional operation having a four cycle branch latency.
This problem can be ameliorated somewhat by employing a technique called speculation or branch prediction to avoid waiting for the result of a comparison. In this technique the processor guesses an outcome for the branch, i.e., whether it is taken and execution jumps or it is not taken and execution proceeds in sequence, and begins executing instructions corresponding to the chosen outcome. Once the true outcome of the conditional operation is known, the results generated by the speculation are confirmed and execution proceeds if the speculative outcome was the correct one, or the results are flushed from the pipeline if the speculation was incorrect.
For example, in the chroma-keying example shown above, when reaching the conditional expression the processor might speculate that the pixel will indeed be blue (since the area of the background is usually larger than that of the foreground subject, this will more often than not be true) and proceed to execute the corresponding branch.
Speculation works well on branches that almost always go one way, e.g., error checks or checks for exceptional conditions, and branches that occur in repeatable patterns, e.g., the return branch at the end of an iterative loop. It does not yield good results on unbiased, highly data-dependent branches and, given completely random data, will guess correctly only 50% of the time (note, however, that this still represents a 50% usage of otherwise dead branch latency cycles).
Another technique designed to work around branch latency effects is predication (sometimes called a select or a masked vector operation in single instruction, multiple data (SIMD) and vector processors), in which instructions from both sides of a branch are executed and, when the actual comparison outcome is known, only the results generated by the correct branch are retained. For example, returning to our chroma-keying example, program execution would proceed to execute instructions for background processing and instructions for foreground processing and, if the pixel in question is found to be blue, the results corresponding to foreground processing would be deleted. Predication is necessarily limited to an efficiency of 50% compared to normal execution, since half the instructions executed will always be incorrect. Further, if comparisons are nested so that more than two outcomes are possible, the maximum efficiency of the technique is correspondingly reduced (of course, the efficiency of speculation also decreases with an increase in possible comparison outcomes).